neuronpedia/dist/index.mjs.map

Neuronpedia (Official)

{"version":3,"sources":["../src/autointerp/score.ts","../src/types/autointerp/score/embedding.ts","../src/types/autointerp/score/fuzz_detection.ts","../src/const.ts","../src/autointerp/explain.ts","../src/types/autointerp/explain/default.ts","../src/types/webapp/api/sae_eval.ts"],"sourcesContent":["import axios from \"axios\";\nimport { postScoreEmbedding } from \"../types/autointerp/score/embedding\";\nimport { postScoreFuzzDetection } from \"../types/autointerp/score/fuzz_detection\";\nimport type { NPActivation } from \"../types/common/activation\";\nimport { AUTOINTERP_SERVER } from \"../const\";\n\nexport async function getEleutherScoreFuzzDetection(\n  type: \"fuzz\" | \"detection\",\n  activations: NPActivation[],\n  explanation: string,\n  openRouterKey: string,\n  openRouterModel: string,\n  inferenceServerSecret: string\n) {\n  axios.defaults.baseURL = AUTOINTERP_SERVER;\n  const result = await postScoreFuzzDetection({\n    type: type,\n    activations: activations,\n    explanation: explanation,\n    openrouter_key: openRouterKey,\n    model: openRouterModel,\n    secret: inferenceServerSecret,\n  });\n  if (result.status !== 200 || !result.data) {\n    throw new Error(\n      \"Failed to generate score: \" + JSON.stringify(result.statusText)\n    );\n  }\n  return result.data;\n}\n\nexport async function getEleutherScoreEmbedding(\n  activations: NPActivation[],\n  explanation: string,\n  inferenceServerSecret: string\n) {\n  axios.defaults.baseURL = AUTOINTERP_SERVER;\n  const result = await postScoreEmbedding({\n    activations: activations,\n    explanation: explanation,\n    secret: inferenceServerSecret,\n  });\n  if (result.status !== 200 || !result.data) {\n    throw new Error(\n      \"Failed to generate score: \" + JSON.stringify(result.statusText)\n    );\n  }\n  return result.data;\n}\n","/**\n * Generated by orval v7.4.1 🍺\n * Do not edit manually.\n * Auto-Interp Server API\n * API for generating and scoring explanations of neuron/feature behavior\n * OpenAPI spec version: 1.0.0\n */\nimport axios from 'axios'\nimport type {\n  AxiosRequestConfig,\n  AxiosResponse\n} from 'axios'\n/**\n * An activation record containing tokens and their corresponding activation values\n */\nexport interface NPActivation {\n  /** List of tokens for this text */\n  tokens: string[];\n  /** Activation values corresponding to each token */\n  values: number[];\n}\n\nexport interface NPScoreEmbeddingResponse {\n  /** The score from 0 to 1 */\n  score: number;\n  /** Detailed breakdown of the embedding outputs */\n  breakdown: NPScoreEmbeddingOutput[];\n}\n\n/**\n * Request model for scoring explanations using embedding similarity\n */\nexport interface NPScoreEmbeddingRequest {\n  /** List of activation records to analyze */\n  activations: NPActivation[];\n  /** The explanation to evaluate */\n  explanation: string;\n  /** Authentication secret for the API */\n  secret: string;\n}\n\n/**\n * Quantile or neighbor distance\n */\nexport type NPScoreEmbeddingOutputDistance = number | number;\n\n/**\n * The \"scorer.__call__\" result's score breakdown. With exception of fixing similarity to change to number instead of array of number, type is copied from https://github.com/EleutherAI/sae-auto-interp/blob/3659ff3bfefbe2628d37484e5bcc0087a5b10a27/sae_auto_interp/scorers/embedding/embedding.py#L20\n */\nexport interface NPScoreEmbeddingOutput {\n  /** The text that was used to evaluate the similarity */\n  text: string;\n  /** Quantile or neighbor distance */\n  distance: NPScoreEmbeddingOutputDistance;\n  /** What is the similarity of the example to the explanation */\n  similarity: number;\n}\n\n\n\n\n\n  /**\n * @summary Score an explanation using embedding similarity, using the dunzhang/stella_en_400M_v5 model.\n */\nexport const postScoreEmbedding = <TData = AxiosResponse<NPScoreEmbeddingResponse>>(\n    nPScoreEmbeddingRequest: NPScoreEmbeddingRequest, options?: AxiosRequestConfig\n ): Promise<TData> => {\n    return axios.post(\n      `/score/embedding`,\n      nPScoreEmbeddingRequest,options\n    );\n  }\n\nexport type PostScoreEmbeddingResult = AxiosResponse<NPScoreEmbeddingResponse>\n","/**\n * Generated by orval v7.4.1 🍺\n * Do not edit manually.\n * Auto-Interp Server API\n * API for generating and scoring explanations of neuron/feature behavior\n * OpenAPI spec version: 1.0.0\n */\nimport axios from 'axios'\nimport type {\n  AxiosRequestConfig,\n  AxiosResponse\n} from 'axios'\n/**\n * An activation record containing tokens and their corresponding activation values\n */\nexport interface NPActivation {\n  /** List of tokens for this text */\n  tokens: string[];\n  /** Activation values corresponding to each token */\n  values: number[];\n}\n\n/**\n * Request model for scoring explanations using fuzzing or detection methods\n */\nexport interface NPScoreFuzzDetectionRequest {\n  /** List of activation records to analyze */\n  activations: NPActivation[];\n  /** The explanation to evaluate */\n  explanation: string;\n  /** API key for OpenRouter service */\n  openrouter_key: string;\n  /** Model identifier to use for scoring */\n  model: string;\n  /** Type of scoring to perform - either fuzzing or detection */\n  type: NPScoreFuzzDetectionType;\n  /** Authentication secret for the API */\n  secret: string;\n}\n\n/**\n * Quantile or neighbor distance\n */\nexport type NPScoreClassifierOutputDistance = number | number;\n\n/**\n * The \"scorer.__call__\" result's score breakdown. Type copied from https://github.com/EleutherAI/sae-auto-interp/blob/3659ff3bfefbe2628d37484e5bcc0087a5b10a27/sae_auto_interp/scorers/classifier/sample.py#L19\n */\nexport interface NPScoreClassifierOutput {\n  /** List of strings */\n  str_tokens?: string[];\n  /** List of floats */\n  activations?: number[];\n  /** Quantile or neighbor distance */\n  distance?: NPScoreClassifierOutputDistance;\n  /** Whether the example is activating or not */\n  ground_truth?: boolean;\n  /** Whether the model predicted the example activating or not */\n  prediction?: boolean;\n  /** Whether the sample is highlighted */\n  highlighted?: boolean;\n  /** The probability of the example activating */\n  probability?: number;\n  /** Whether the prediction is correct */\n  correct?: boolean;\n}\n\nexport interface NPScoreFuzzDetectionResponse {\n  /** The score from 0 to 1 */\n  score: number;\n  /** Detailed breakdown of the classification outputs */\n  breakdown: NPScoreClassifierOutput[];\n}\n\n/**\n * Type of scoring method, either fuzz or detection\n */\nexport type NPScoreFuzzDetectionType = typeof NPScoreFuzzDetectionType[keyof typeof NPScoreFuzzDetectionType];\n\n\n// eslint-disable-next-line @typescript-eslint/no-redeclare\nexport const NPScoreFuzzDetectionType = {\n  fuzz: 'fuzz',\n  detection: 'detection',\n} as const;\n\n\n\n\n\n  /**\n * @summary Score an explanation using fuzzing or detection methods\n */\nexport const postScoreFuzzDetection = <TData = AxiosResponse<NPScoreFuzzDetectionResponse>>(\n    nPScoreFuzzDetectionRequest: NPScoreFuzzDetectionRequest, options?: AxiosRequestConfig\n ): Promise<TData> => {\n    return axios.post(\n      `/score/fuzz_detection`,\n      nPScoreFuzzDetectionRequest,options\n    );\n  }\n\nexport type PostScoreFuzzDetectionResult = AxiosResponse<NPScoreFuzzDetectionResponse>\n","// these should be defined in the environment of the app importing this package\nconst USE_LOCAL_AUTOINTERP = process.env.USE_LOCAL_AUTOINTERP === \"true\";\nconst AUTOINTERP_SERVER = USE_LOCAL_AUTOINTERP\n  ? \"http://127.0.0.1:5003\"\n  : process.env.AUTOINTERP_SERVER;\n\nexport { AUTOINTERP_SERVER };\n","import axios from \"axios\";\nimport { postExplainDefault } from \"../types/autointerp/explain/default\";\nimport type { NPActivation } from \"../types/common/activation\";\nimport { AUTOINTERP_SERVER } from \"../const\";\n\nexport async function getEleutherExplanationDefault(\n  activations: NPActivation[],\n  openRouterKey: string,\n  openRouterModel: string,\n  inferenceServerSecret: string\n) {\n  axios.defaults.baseURL = AUTOINTERP_SERVER;\n  const result = await postExplainDefault({\n    activations: activations,\n    openrouter_key: openRouterKey,\n    model: openRouterModel,\n    secret: inferenceServerSecret,\n  });\n\n  if (result.status !== 200 || !result.data) {\n    throw new Error(\n      \"Failed to generate explanation: \" + JSON.stringify(result.statusText)\n    );\n  }\n  return result.data;\n}\n","/**\n * Generated by orval v7.4.1 🍺\n * Do not edit manually.\n * Auto-Interp Server API - Explanation Generation\n * Generating an explanation via OpenRouter using the default method\n * OpenAPI spec version: 1.0.0\n */\nimport axios from 'axios'\nimport type {\n  AxiosRequestConfig,\n  AxiosResponse\n} from 'axios'\n/**\n * An activation record containing tokens and their corresponding activation values\n */\nexport interface NPActivation {\n  /** List of tokens for this text */\n  tokens: string[];\n  /** Activation values corresponding to each token */\n  values: number[];\n}\n\nexport interface NPExplainDefaultResponse {\n  /** The generated explanation for the given set of activations */\n  explanation: string;\n}\n\n/**\n * Request model for generating explanations of neuron/feature behavior\n */\nexport interface NPExplainDefaultRequest {\n  /** List of activation records to analyze */\n  activations: NPActivation[];\n  /** API key for OpenRouter service */\n  openrouter_key: string;\n  /** Model identifier to use for explanation generation */\n  model: string;\n  /** Authentication secret for the API */\n  secret: string;\n}\n\n\n\n\n\n  /**\n * @summary Generate an explanation for neuron/feature behavior using the default explainer\n */\nexport const postExplainDefault = <TData = AxiosResponse<NPExplainDefaultResponse>>(\n    nPExplainDefaultRequest: NPExplainDefaultRequest, options?: AxiosRequestConfig\n ): Promise<TData> => {\n    return axios.post(\n      `/explain/default`,\n      nPExplainDefaultRequest,options\n    );\n  }\n\nexport type PostExplainDefaultResult = AxiosResponse<NPExplainDefaultResponse>\n","/**\n * Generated by orval v7.4.1 🍺\n * Do not edit manually.\n * Fetch SAE eval results\n * OpenAPI spec version: 1.0.0\n */\nimport axios from 'axios'\nimport type {\n  AxiosRequestConfig,\n  AxiosResponse\n} from 'axios'\nexport interface NPSaeEvalListResponse {\n  /** List of SAE eval types and their results */\n  evalTypes?: NPSaeEvalType[];\n}\n\nexport interface NPSaeEvalListRequest { [key: string]: unknown }\n\n/**\n * An SAE eval result containing the outputs (metrics, config, etc) of the eval\n */\nexport interface NPSaeEval {\n  /** ID of the SAE Eval */\n  id: string;\n  /** SAE Eval Type - should equal the \"name\" property of NPSaeEvalType */\n  typeName: string;\n  /** The model ID of the SAE being evaluated */\n  modelId: string;\n  /** The source ID of the SAE being evaluated */\n  sourceId: string;\n  /** The output data, matching the outputSchema of the NPSaeEvalType, with the exception of eval_result_details, which will go under detailedMetrics */\n  output: string;\n  /** The eval_result_details of the output data. Kept in a separate field/column due to large size. */\n  detailedMetrics?: string;\n}\n\n/**\n * A type of SAE Eval (eg absorption, sparse probing, etc). Can contain an array of the actual eval results.\n */\nexport interface NPSaeEvalType {\n  /** ID of the Eval Type */\n  name: string;\n  /** A UI-friendly display name of the eval type */\n  displayName: string;\n  /** A brief description of the eval */\n  description: string;\n  /** JSON output schema used to parse the eval output results. From the SAEBench repo - for example https://github.com/adamkarvonen/SAEBench/blob/main/sae_bench/evals/autointerp/eval_output_schema_autointerp.json */\n  outputSchema: string;\n  /** URL to the eval */\n  url?: string;\n  /** The actual eval results of this type */\n  evals?: NPSaeEval[];\n}\n\n\n\n\n\n  /**\n * @summary List all SAE Evals\n */\nexport const postApiSaeEval = <TData = AxiosResponse<NPSaeEvalListResponse>>(\n    nPSaeEvalListRequest: NPSaeEvalListRequest, options?: AxiosRequestConfig\n ): Promise<TData> => {\n    return axios.post(\n      `/api/sae-eval`,\n      nPSaeEvalListRequest,options\n    );\n  }\n\nexport type PostApiSaeEvalResult = AxiosResponse<NPSaeEvalListResponse>\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;AAAA,OAAOA,YAAW;;;ACOlB,OAAO,WAAW;AA0DX,IAAM,qBAAqB,CAC9B,yBAAkD,YAChC;AAClB,SAAO,MAAM;AAAA,IACX;AAAA,IACA;AAAA,IAAwB;AAAA,EAC1B;AACF;;;ACjEF,OAAOC,YAAW;AA0EX,IAAM,2BAA2B;AAAA,EACtC,MAAM;AAAA,EACN,WAAW;AACb;AASO,IAAM,yBAAyB,CAClC,6BAA0D,YACxC;AAClB,SAAOA,OAAM;AAAA,IACX;AAAA,IACA;AAAA,IAA4B;AAAA,EAC9B;AACF;;;ACnGF,IAAM,uBAAuB,QAAQ,IAAI,yBAAyB;AAClE,IAAM,oBAAoB,uBACtB,0BACA,QAAQ,IAAI;;;AHEhB,SAAsB,8BACpB,MACA,aACA,aACA,eACA,iBACA,uBACA;AAAA;AACA,IAAAC,OAAM,SAAS,UAAU;AACzB,UAAM,SAAS,MAAM,uBAAuB;AAAA,MAC1C;AAAA,MACA;AAAA,MACA;AAAA,MACA,gBAAgB;AAAA,MAChB,OAAO;AAAA,MACP,QAAQ;AAAA,IACV,CAAC;AACD,QAAI,OAAO,WAAW,OAAO,CAAC,OAAO,MAAM;AACzC,YAAM,IAAI;AAAA,QACR,+BAA+B,KAAK,UAAU,OAAO,UAAU;AAAA,MACjE;AAAA,IACF;AACA,WAAO,OAAO;AAAA,EAChB;AAAA;AAEA,SAAsB,0BACpB,aACA,aACA,uBACA;AAAA;AACA,IAAAA,OAAM,SAAS,UAAU;AACzB,UAAM,SAAS,MAAM,mBAAmB;AAAA,MACtC;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,IACV,CAAC;AACD,QAAI,OAAO,WAAW,OAAO,CAAC,OAAO,MAAM;AACzC,YAAM,IAAI;AAAA,QACR,+BAA+B,KAAK,UAAU,OAAO,UAAU;AAAA,MACjE;AAAA,IACF;AACA,WAAO,OAAO;AAAA,EAChB;AAAA;;;AIhDA,OAAOC,YAAW;;;ACOlB,OAAOC,YAAW;AAyCX,IAAM,qBAAqB,CAC9B,yBAAkD,YAChC;AAClB,SAAOA,OAAM;AAAA,IACX;AAAA,IACA;AAAA,IAAwB;AAAA,EAC1B;AACF;;;ADlDF,SAAsB,8BACpB,aACA,eACA,iBACA,uBACA;AAAA;AACA,IAAAC,OAAM,SAAS,UAAU;AACzB,UAAM,SAAS,MAAM,mBAAmB;AAAA,MACtC;AAAA,MACA,gBAAgB;AAAA,MAChB,OAAO;AAAA,MACP,QAAQ;AAAA,IACV,CAAC;AAED,QAAI,OAAO,WAAW,OAAO,CAAC,OAAO,MAAM;AACzC,YAAM,IAAI;AAAA,QACR,qCAAqC,KAAK,UAAU,OAAO,UAAU;AAAA,MACvE;AAAA,IACF;AACA,WAAO,OAAO;AAAA,EAChB;AAAA;;;AEnBA,OAAOC,YAAW;AAuDX,IAAM,iBAAiB,CAC1B,sBAA4C,YAC1B;AAClB,SAAOA,OAAM;AAAA,IACX;AAAA,IACA;AAAA,IAAqB;AAAA,EACvB;AACF;","names":["axios","axios","axios","axios","axios","axios","axios"]}